- F5 research finds 78% of enterprises run AI inference, creating distributed system complexity
- Telcos are positioned to evolve beyond connectivity into AI infrastructure hosting, offering compute, routing and sovereign cloud services
- Identity and intent — what F5 calls "layer eight" — are emerging as the next AI security frontier, beyond what today's tools handle
Enterprises have gone beyond AI pilots and proofs of concept and are deploying AI at scale, creating management complexity greater than most organizations anticipated.
F5's 2026 State of Application Strategy report finds that organizations face challenges routing traffic between models based on application requirements, securing traffic and governing AI in production, Shawn Wormke, F5 SVP of product management, told Fierce.
The report, based on responses from more than 1,100 IT decision-makers globally, found that 78% of organizations now run their own AI inference rather than relying on public AI services, and that enterprises are operating an average of seven AI models simultaneously in production.
The effect is to transform AI into a distributed systems problem. As with hybrid multi-cloud application delivery — a challenge enterprises and their service providers have been navigating for years — managing inference across multiple models demands routing, fallback logic, security, observability and governance at scale. The report found that 52% of organizations are already chaining or orchestrating multiple models, and that multi-model orchestration has outpaced even prompt engineering as the leading adaptation technique, suggesting enterprises are designing systems rather than tuning individual models.
What this means for telcos
For communications service providers, the implications cut both operationally and commercially.
On the operational side, inference is following the same trajectory as general application workloads — spreading from centralized data centers toward edges, devices and distributed environments. As agentic AI frameworks proliferate across multiple systems, the volume and latency sensitivity of AI traffic will place new demands on network infrastructure, Wormke said.
"That puts a ton of pressure on telcos," he said. "The amount of traffic will rise a lot, and they're going to have to be in the game of routing that traffic, making sure that as latency becomes more of a concern, they're able to deliver on those SLAs."
That's a big challenge, but it's also a big opportunity. Telcos have long struggled to move beyond the dumb-pipe role, but AI infrastructure may offer a credible path forward.
Service providers are naturally positioned to compete on compute and hosting demands of running inference at scale — access to specialized hardware, low-latency delivery at the edge and sovereignty requirements, Wormke said.
"Service providers are going to have to get away from just being dumb pipes and be able to provide higher-level services to their enterprise customers," he said. He noted that sovereign cloud deployments are already accelerating in Europe and Asia, and that neoclouds are emerging directly from telco companies in Asia as a result.
Layer eight: context, intent and the limits of today's security
The deepest challenge ahead, Wormke said, is what F5 internally calls "layer eight" — context and intent.
Layer seven of the OSI model covers what happens at the application protocol layer. Layer eight addresses the non-deterministic nature of AI: the need to understand the original intent of an agentic action and detect when that intent has drifted during execution.
"You need to understand the original intent of what that agentic action was supposed to do, understand the drift from that and be able to make security decisions based on that original intent versus what the outcome actually is," Wormke said.
Because AI systems are non-deterministic, evaluating individual transactions in isolation is insufficient. Layer eight requires assembling context across an entire interaction — agent identity, permissions, the original instruction and the actual outcome — to render a defensible security judgment. That is not a problem today's security tooling is built to solve.